Blog
Engineering and product notes from the team behind SpeechButton.
I built a 20ms voice layer for AI agents on Mac — here's the architecture
The latency budget line by line, why the Apple Neural Engine beats the GPU for a tool used in five-second bursts, why the core was rewritten from Swift to Rust, and how chunk-on-pause streaming works.
Read →Three hotkeys, thirty agents — how I route my voice across an AI fleet
Thirty agents across a laptop, a remote Linux box and deployed HTTP endpoints, driven by three hotkeys. The hotkey is the channel: cursor, SSH session, Linear issue, deployed agent — with a local LLM shaping the payload in between.
Read →I paid for the best-rated Mac dictation app — two days later I started writing my own
A 300ms delay that ate the first word, a hands-free mode that didn't stream, and a single output target. The tool wasn't bad — it was built for a different person.
Read →Only your voice — Voice Fingerprint in SpeechButton 2.12
SpeechButton 2.12 transcribes the person who set it up — and ignores everyone else. Background voices, family in the next room, audio on speakers: filtered out before they reach the model. 100% on device.
Read →